Sonar scene analysis using neurobionic sound segregation

Abstract

A computing architecture is being produced that automates primitive and schemea-based streaming of sounds and thereby achieves better real-time, in-situ analyses of complicated sonar scenes. The computational models are called the neural beamformers (NBFs). A brief qualitative overview of three beamformers is given: the crossbar beamformer is based on the Hopfield crossbar circuit; the multivector beamformer is related to Kohonen feature map learning; and the neurobionic beamformer is really a network of beamformers and combines elements of the other two beamformers. In experiments using an array of microphones operated in a laboratory room, an NBF was able to locate a sound source while exhibiting tolerance to sounds arriving at the array via a reflected path once the processing had seen the onset of the direct path excitation from the source.<>